imageSVD

Utility for comressing/decompressing images using SVD.

This utility takes a BMP image as input and compresses it using SVD for each of the RGB channels. Examples of using the utility, as well as the structure of compressed images are presented below.

Usage

cdisvd compress --out-file compressed --method power-iterations input-file.bmp
cdisvd decompress --out-file decompressed.bmp input-file

Commads and arguments

cdisvd --help
cdisvd compress --help
cdisvd decompress --help

Available methods

• numpy:

numpy.linalg.svd()

• power-iterations: http://www.cs.yale.edu/homes/el327/datamining2013aFiles/07_singular_value_decomposition.pdf
• block-power-iter: https://www.degruyter.com/document/doi/10.1515/jisys-2018-0034/html

Compressed format

The compressed image has the following structure.

Part	Size bytes	Desc
Header	16	Contains general information about the compressed image.
Body	Variable	Contains an image using matrices.

Header

Offset hex	Size bytes	Name	Type	Desc
00	4	signature	uint32	Signature of compressed file, must be "BOOB" in ASCII.
04	4	Width	uint32	Width of image in pixels.
08	4	Height	uint32	Height of image in pixels.
0C	4	K	uint32	One side size of matrix, computes automaticly.

Body

There are tree part for each RBG's channel. All values are written sequentially, without any alignments or separators.

Name	Size bytes	Type	Description
U	4 * width * k	float32	Matrix containing of left singular vectors.
S	4 * k	float32	Vector containing singular numbers
Vt	4 * height * k	float32	Matrix containing of right singular vector.